一种基于跨领域典型相关性分析的迁移学习方法

摘要

As one of the most important research directions of transfer learning,feature-representation-transfer approaches focus on the correlation between bridge features and all the other specific features from different domains and reduce the difference between the domains by learning some relevant features,have attracted wide attention and study.Canonical correlation analysis (CCA)is a statistical analysis tool,used to analyze the correlation between the two sets of random variables.By introducing CCA to transfer learning,this paper developed a canonical correlation analysis across different domains called CCADD(Canonical Correlation Analysis across Different Domains),followed by the idea of feature-representation-transfer approaches.Under the premise of maintaining the correlation between bridge features across all domains and specific features from different domains respectively,this algorithm selects an appropriate combination of basis vectors to train the classifier,in which the projected relevant features have similar discrimi-nation.Experimental results on the 864 classification problems in 20Newsgroups,as well as 12 classification problems in multi-domain sentiment analysis datasets,show that CCADD can significantly improve the cross domain prediction accuracy of a baseline non-transfer method.%作为迁移学习的一个重要研究方向，基于特征映射的方法学习各领域特有特征与领域共享特征之间的相关性，通过一些相关特征减少领域之间的差异，已经获得了广泛的关注和研究。典型相关性分析是一种用来分析两组随机变量之间相关性的统计分析工具。将典型相关性分析引入迁移学习，结合基于特征映射迁移学习的思路，提出了一种跨领域典型相关性分析算法。该算法在保持各领域特有特征与领域共享特征相关性的基础上，通过选择合适的基向量组合训练分类器，使降维后的相关特征在领域间具有相似的判别性。在20Newsgroups 上864个分类问题以及多领域情感分析数据集上12个分类问题的实验结果表明，跨领域典型相关性分析算法可以有效地提高跨领域迁移分类准确率。